Non-convex regularized robust multimodal feature selection via self-representation learning for Alzheimer’s disease diagnosis

Abstract

Multimodal neuroimaging data fusion has become a key research direction in Alzheimer’s Disease (AD) diagnosis. However, existing methods face challenges such as (1) Limited robustness against outliers and noise, which hampers effective feature selection; (2) Limitations of conventional convex approximation methods, such as the ${\ell }_{2,1}$ norm, in approximating the ideal ${\ell }_{2,0}$ norm, making it challenging to capture sparse structures accurately; (3) Inadequate modeling of feature correlations, leading to missed identification of synergistic feature groups. To address these issues, this study proposes a Non-Convex Regularized Robust Multimodal Feature Selection method via Self-Representation Learning for Alzheimer’s Disease diagnosis (NCRRFS). Specifically, self-representation learning is employed to model the error terms of anomalous samples, enabling the adaptive detection and correction of abnormal data, thereby enhancing the robustness of the model. Furthermore, an ${\ell }_{2,\gamma }$ norm row sparsity constraint based on the Smoothly Clipped Absolute Deviation (SCAD) function is designed to more accurately approximate the ${\ell }_{2,0}$ norm. Additionally, a graph-structured regularization based on Pearson correlation promotes the selection of synergistic feature groups. Extensive experimental results demonstrate the effectiveness and superiority of the proposed method in the Alzheimer’s disease classification task.